Comment by carapace | Hacker News Reader

Comment by carapace | Hacker News Reader

I'm a fan of self-driving vehicles (can we please call them auto-autos?) because they are, as this article points out, already safer than human drivers.

Still, it is and always has been completely and totally insane to test these heavy fast robots on public streets. I never in a million years would have thought we would do that.

The obvious thing to do is:

A) Start with light slow robots, like a golf cart covered in foam, and work your way up to Knight Industries Two Thousand. Going directly for KITT is a fetish.

B) Test it in one of those artificial cities with, you know, people who have at least signed waivers to be human test dummies for your two-ton high-speed killer robots?

E.g.:

> Mcity is a 32-acre simulated urban and suburban environment that includes a network of roads with intersections, traffic signs and signals, streetlights, building facades, sidewalks and construction obstacles. It is designed to support rigorous, repeatable testing of new technologies before they are tried out on public streets and highways.

https://news.umich.edu/u-m-opens-mcity-test-environment-for-...

You have described more or less exactly how Waymo did their rollout.

A. They used to have these cute little cars (e.g. golf cart sized) that couldn't go over 30 mph, long before they rolled out the Chryslers and Jaguars

B. My memory is fuzzy on the details, but I'm sure there's public info on this: Waymo owned or leased some massive plot of land where they did exactly that, testing all sorts of chaotic scenarios (weird construction, skateboards coming out of nowhere, etc)

That's great! It's probably why they haven't killed anybody yet.

The other thing that should be done is to create a uniform federal test that computer driven cars must pass before being allowed on the streets. The test would have a large variety of environments (fast, slow, construction zones, farmers markets) as well as conditions (wet, dry, light, dark). The test would also measure how the computer driven car reacts when one or all of its sensors are damaged. Only after passing the test should a car be allowed on the road. Any software or hardware changes would necessitate a full re-test.

None of this "well the manufacturer says it works so let's just let 'er rip" that has already gotten people killed.

And yes human drivers should have to pass the same test in the vehicle they intend to drive.

> a uniform federal test

Just like the existing uniform state run test that shows how all licensed human drivers are safe? It shouldn't need to be pointed out how any kind of standardized test isn't going to prove or even realistically evaluate the safety of a system like this. We can't even get a standardized test for GPU performance that manufacturers don't cheat on.

> The test would also measure how the computer driven car reacts when one or all of its sensors are damaged.

I do agree with this, though: we already expect cars to behave well in the face of hardware failures. Autonomous or no, cars should have requirements for failure modes.

It's unfortunate that the National Electrical Code doesn't exist. In fact, it is unfortunate that we can't possibly have national regulations for anything at all. If only this were possible.

> Still, it is and always has been completely and totally insane to test these heavy fast robots on public streets

Another comment pointed out that Google did roll out in an extremely safe way, but to this point: do you still consider this to be the testing phase? You might call it that but if we're at a point where "they are, as this article points out, already safer than human drivers" (and in particular, nearly 2^3 times better in some cases) that would suggest that the fundamental concept of the vehicles no longer need testing.

It's simply not possible to develop autonomous vehicles completely in simulated environments. There's two kinds of testing:

1. Testing that the car behaves appropriately such that it's suitable for public use. This is the phase where you're not sure that it's safe.

2. Iteratively improving the vehicle after it's been deemed safe. This is where you're rolling out changes that you believe will improve safety and comfort on an existing safe product.

Google already did #1. They did it for over a decade. If you believe that #1 and #2 are not two separate phases, then if #1 can't be evaluated with real human drivers you can't ever get to a place where you deem the vehicles safe. There's no threshold where you can confidently say the cars will behave well because you've never put them in front of real, unpaid humans on real, non-simulated roads. How could you call them safe in good conscience if you've never tested them outside of a test environment? How do you know you've controlled every last variable? You simply don't and can't.

> can we please call them auto-autos?

Proposal accepted.

Thank you. :)

Literally 'self-selfs' in some languages. 'Automobile' is a clue.

In English too. :)

(The first "auto-" is self-moving as contrasted with the cart that needed an external motive (horse etc.), then the second "auto-" is self-directing with sensors and feedback systems forming a crude "mind" with a "self image" (all efficient regulators contain a model of themselves[1]). A self-motile and self-guided machine is an auto-auto-mobile. I love it.)

[1] https://en.wikipedia.org/wiki/Good_Regulator

Reply on news.ycombinator.com